Optimization EfficientNetV2 model variant using Grad-CAM for multiple MRI brain tumor classification

Denisa Septalian Alhamda (university of Pembangunan Nasional Veteran Jawa Timur)
Wahyu Syaifullah J (university of Pembangunan Nasional Veteran Jawa Timur)
Prasetyaning Estu Pratiwi (university of Pembangunan Nasional Veteran Jawa Timur)
Surjo Hadi (university of Yos Soedarso Surabaya)
Wan Suryani Wan Awang (University Sultan Zainal Abidin Besut Campus, 22200 Besut, Terengganu, Malaysia)
I Gede Susrama Mas Diyasa (UPN "Veteran" Jawa Timur)

Jurnal Ilmiah Kursor
Vol. 13 No. 3 (2026)

Fast and accurate diagnosis plays a critical role in effectively treating brain tumors. This study optimized and evaluated the EfficientNetV2 architecture through transfer learning, fine-tuning, and data augmentation, using three variants Small, Medium, and Large to classify MRI images into four categories: glioma, meningioma, pituitary tumors, and no tumor. Grad-CAM visualization was employed to enhance interpretability, providing a clear view of the critical regions in the MRI images that influenced the model’s decisions. Grad-CAM was tested across all model variants, and the best results were observed with EfficientNetV2-Large, where the model successfully highlighted the key areas associated with brain tumors. Among the variants, EfficientNetV2-Large achieved the best performance, with 99.85% accuracy, 99.60% precision, 99.65% recall, and 99.50% F1-score. However, this model required the longest computation time of 288 seconds per step, which may not be feasible in resource- constrained environments. Overall, this study underscores the potential of EfficientNetV2 models in revolutionizing brain tumor diagnosis by balancing accuracy, efficiency, and interpretability through advanced optimization techniques.Key words: Brain tumors, MRI classification, EfficientNetV2, Grad-CAM, Deep learning.